Zack Shapiro is a New York lawyer who built a two-person law firm on Claude and pulled in roughly 7.5 million views telling the story. In his essay The Two Clocks he lands on an idea that explains why AI stalls inside most companies. Picture two clocks. One tracks how fast model capabilities improve. The other tracks how fast organizations learn to use them. The first sprints, the second crawls, and the business of the decade gets built on that gap.
We ship AI agents and automation every week, so we read this less as a nice metaphor and more as a description of our job. The gap is real, it is measurable, and what closes it is not your choice of model but your AI implementation strategy: how you redesign the work, who owns it, and where you start. Below are the numbers on both clocks and what we have learned across 100+ projects.
The two-clocks metaphor: capability and institutions tick at different speeds
The two clocks are the speed at which models get cheaper and smarter, and the speed at which companies rewire the work around them. The first outruns the second by years.
On the first clock the numbers are stark. The cost of intelligence is falling faster than anything in the history of computing. Inference matching GPT-3.5 dropped from about $20 per million tokens in late 2022 to $0.07 by late 2024, a 280x collapse in two years (Silicon Canals). Frontier models get roughly 10x cheaper a year, and a new state-of-the-art release lands on average every 90 days (TokenMix).
For a business that means one thing: a capability that simply did not exist a year ago, an agent that runs a conversation or reads inbound documents, now costs pennies per call. The technical ceiling rises every quarter.
The second clock runs on a different calendar. A pilot, a security review, an integration, and training people take quarters, sometimes years. And as Shapiro points out, the gap will not close on its own, because the clocks are built to run at different speeds. Whoever turns capability jumps into internal change faster than their competitors takes the market.
The gap in numbers: 95% of pilots with no ROI and "pilot purgatory"
While AI capability accelerates, the payoff inside companies barely moves. According to MIT, about 95% of enterprise GenAI pilots deliver no measurable impact on profit.
Here is what the second clock looks like across 2025 research:
| What was measured | Figure | Source |
|---|---|---|
| GenAI pilots with no measurable P&L impact | ~95% | MIT, "The GenAI Divide" |
| Companies using AI in at least one function | 88% | McKinsey State of AI 2025 |
| Of those, scaling AI across the enterprise | ~1/3 | McKinsey State of AI 2025 |
| "High performers" with >5% EBIT impact | ~5.5% | McKinsey State of AI 2025 |
| Abandoned most AI initiatives in 2025 | 42% (up from 17%) | S&P Global |
| Getting no material value from AI | ~60% | BCG |
MIT's report, "The GenAI Divide: State of AI in Business 2025," is blunt: the problem is not model quality but a learning gap. General tools like ChatGPT shine in demos yet stall in real workflows because they do not adapt to them. Meanwhile enterprise GenAI spend has already hit $30–40 billion (Fortune). One more detail from the same report: more than half of GenAI budgets go to sales and marketing, yet the fastest payback shows up in the back office, where teams automate routine work and cut manual operations. Companies fund the storefront while the money sits in the warehouse.
McKinsey surveyed 1,993 leaders across 105 countries and found the same shape: 88% have AI in at least one function, but only about a third scaled it, and roughly 5.5% became high performers (McKinsey). The rest sit in what people now call pilot purgatory. The distance between 88% and one third is exactly the chasm Shapiro describes: almost everyone tried, only a few turned it into money. BCG adds that around 60% of companies see no material value from AI at all (BCG). And per S&P Global, 42% of companies abandoned most of their AI initiatives in 2025, up from 17% a year earlier.
The organization stalls, not the model
The main barrier to AI adoption is not model quality but how the work around it is built. Google Cloud's DORA research estimates that 70% of AI transformation value comes from people and process, 20% from infrastructure, and only 10% from the algorithms.
The rest of the numbers point the same way. McKinsey's November 2025 report finds that nearly 80% of organizations layer AI on top of existing processes without redesigning them, which produces modest gains that get eaten by manual workarounds in the unchanged process. Deloitte's survey of 3,235 leaders across 24 countries found that only 37% invested seriously in change management, training, and incentives (Deloitte). And one more detail from McKinsey: the factor most correlated with financial impact is not budget size but the CEO personally owning AI governance.
Add data to that: a model will not pull meaning from tables nobody cleaned, and it will not invent a process that was never written down. Add trust: people will not hand a task to an agent they do not trust, so a pilot with no clear "why does this help me" dies quietly.
This matches what we see on projects. Half of our clients arrive not with "which model" but with "I do not understand where to start in our case." We wrote up those patterns in what we learned after 100+ AI consultations: the model was almost never the bottleneck. The bottleneck was the process, the data, and who inside the company owned the result.
The "McKinsey of the AI era": who actually closes the gap
The "McKinsey of the AI era" is not a consultancy selling strategy in slides. It is the team that gets AI into production inside a company's real processes. Most polished decks break on that verb, "gets."
The money is already flowing toward advice. The AI consulting market is valued at roughly $7–11 billion in 2025, growing toward tens of billions by 2030 (Market reports). Accenture alone reported $5.9 billion in GenAI bookings for fiscal 2025. But the gap between the two clocks is closed by implementation, not advice. Classic consulting ends exactly where the hard part starts: CRM and ERP integration, cleaning data, escalation logic to a human, and running the thing in production. Boutique implementation teams win precisely where the Big Four are slow: a short cycle, direct access to the engineers, and accountability for a working result in production rather than for the thickness of the report.
Our angle is simple: we ship working systems, not slides. That is our AI implementation services in one line. For example, 2nd place out of 350+ teams at the Agentic Legal RAG Challenge was not a presentation about RAG but a built, tested document-search system. More examples live in the Gless case studies.
What this means for your company: how to build an AI implementation strategy
A working AI implementation strategy starts with the process, the owner, and the metric, not the model. A narrow pilot with a clear impact on profit beats a sprawling "AI platform" that nobody can operate later.
Five steps that fall straight out of the numbers above:
- Start with the process, not the model. Pick one process with clear rules and money on the line: document parsing, lead qualification, request handling.
- Assign an owner. Top-level ownership correlates most with impact. Without an owner, the initiative drifts into pilot purgatory.
- Redesign the workflow instead of layering AI on top. 80% of companies do the reverse and get gains that manual workarounds eat.
- Invest in people. Only 37% of companies do, and that is where 70% of the value hides.
- Measure P&L, not "we deployed AI." That metric is where the 95% of failed pilots drop out.
In practice it is mundane. Take one process, say inbound document handling, count what it costs in person-hours, build a narrow pilot, measure the result in money, and only then decide whether to scale. Boring, and it lands you in that 5% instead of the 95%.
And keep people in the loop. AI amplifies whoever understands the task rather than replacing them outright, which is the point we made in why AI doesn't replace engineers.
The capability clock will only run faster, and there is nothing you can do about that. The one thing in your control is the second clock: how fast your company turns new capability into working process. If you'd like to scope where to start and how to stay out of that 95%, get in touch.